| Benchmark | DeepSeek-V3.2 | DeepSeek-V4-Flash |
|---|---|---|
| AA Agentic Index | 39.8 | 61.3 |
| AA Intelligence | 32.8 | 51.8 |
| AA-LCR | 70.7 | 74.3 |
| AA-Omniscience | -22.5 | -14.3 |
| aime_2026 | 95.1 | 95.8 |
| arc_agi_1 | 57 | 84 |
| ARC-AGI-2 | 4 | 46 |
| arena_text_factuality | 1433 | 1434 |
| Artificial Analysis Coding Index | 44.2 | 69.1 |
| browsecomp | 51.4 | 73.2 |
| BrowseComp (context management) | 67.6 | 73.2 |
| coding_arena_elo | 1368 | 1431 |
| critpt | 2.9 | 16.6 |
| cybergym | 17.3 | 76.7 |
| gdpval | 18.8 | 52.9 |
| GPQA Diamond | 84 | 90.8 |
| HLE | 40.8 | 38.6 |
| HLE (with tools) | 40.8 | 45.1 |
| hmmt_feb_2026 | 84.1 | 94.8 |
| IFBench | 61 | 79.2 |
| imo_answer_bench | 78.3 | 88.4 |
| LiveCodeBench | 86 | 91.6 |
| MCP Atlas | 62.2 | 69 |
| MMLU-Pro | 86 | 86.4 |
| OmniScience Accuracy | 33 | 40.4 |
| OmniScience Non-Hallucination | 17.3 | 15.6 |
| scicode | 39 | 49.9 |
| simpleqa_verified | 27.5 | 45 |
| SWE-bench Multilingual | 70.2 | 73.3 |
| SWE-bench Pro | 15.6 | 52.6 |
| SWE-bench Verified | 73.1 | 79 |
| TauBench V3 - Banking | 18.8 | 39.4 |
| Terminal-Bench 2.0 | 46.4 | 56.9 |
| Terminal-Bench 2.1 | 46.8 | 82.7 |
| Terminal-Bench Hard | 35.6 | 38.6 |
| toolathlon | 35.2 | 70.3 |
| τ²-Bench Telecom (AA run) | 90.6 | 95.6 |
| τ³-Bench | 69.2 | 22.9 |
Best tracked score per model per benchmark (default configuration; source-attributed). ↓ marks lower-is-better metrics. Open either model for its full surface, provenance and pricing.